Learning policies from offline datasets through offline reinforcement learning (RL) holds promise for scaling data-driven decision-making and avoiding unsafe and costly online interactions. However, real-world data collected from sensors or humans often contains noise and errors, posing a significant challenge for existing offline RL methods. Our study indicates that traditional offline RL methods based on temporal difference learning tend to underperform Decision Transformer (DT) under data corruption, especially when the amount of data is limited. This suggests the potential of sequential modeling for tackling data corruption in offline RL. To further unleash the potential of sequence modeling methods, we propose Robust Decision Transformer (RDT) by incorporating several robust techniques. Specifically, we introduce Gaussian weighted learning and iterative data correction to reduce the effect of corrupted data. Additionally, we leverage embedding dropout to enhance the model's resistance to erroneous inputs. Extensive experiments on MoJoCo, KitChen, and Adroit tasks demonstrate RDT's superior performance under diverse data corruption compared to previous methods. Moreover, RDT exhibits remarkable robustness in a challenging setting that combines training-time data corruption with testing-time observation perturbations. These results highlight the potential of robust sequence modeling for learning from noisy or corrupted offline datasets, thereby promoting the reliable application of offline RL in real-world tasks.

Q-learning played a foundational role in the field reinforcement learning (RL). However, TD algorithms with off-policy data, such as Q-learning, or nonlinear function approximation like deep neural networks require several additional tricks to stabilise training, primarily a replay buffer and target networks. Unfortunately, the delayed updating of frozen network parameters in the target network harms the sample efficiency and, similarly, the replay buffer introduces memory and implementation overheads. In this paper, we investigate whether it is possible to accelerate and simplify TD training while maintaining its stability. Our key theoretical result demonstrates for the first time that regularisation techniques such as LayerNorm can yield provably convergent TD algorithms without the need for a target network, even with off-policy data. Empirically, we find that online, parallelised sampling enabled by vectorised environments stabilises training without the need of a replay buffer. Motivated by these findings, we propose PQN, our simplified deep online Q-Learning algorithm. Surprisingly, this simple algorithm is competitive with more complex methods like: Rainbow in Atari, R2D2 in Hanabi, QMix in Smax, PPO-RNN in Craftax, and can be up to 50x faster than traditional DQN without sacrificing sample efficiency. In an era where PPO has become the go-to RL algorithm, PQN reestablishes Q-learning as a viable alternative. We make our code available at: https://github.com/mttga/purejaxql.

This paper presents a novel approach for unsupervised video summarization using reinforcement learning. It aims to address the existing limitations of current unsupervised methods, including unstable training of adversarial generator-discriminator architectures and reliance on hand-crafted reward functions for quality evaluation. The proposed method is based on the concept that a concise and informative summary should result in a reconstructed video that closely resembles the original. The summarizer model assigns an importance score to each frame and generates a video summary. In the proposed scheme, reinforcement learning, coupled with a unique reward generation pipeline, is employed to train the summarizer model. The reward generation pipeline trains the summarizer to create summaries that lead to improved reconstructions. It comprises a generator model capable of reconstructing masked frames from a partially masked video, along with a reward mechanism that compares the reconstructed video from the summary against the original. The video generator is trained in a self-supervised manner to reconstruct randomly masked frames, enhancing its ability to generate accurate summaries. This training pipeline results in a summarizer model that better mimics human-generated video summaries compared to methods relying on hand-crafted rewards. The training process consists of two stable and isolated training steps, unlike adversarial architectures. Experimental results demonstrate promising performance, with F-scores of 62.3 and 54.5 on TVSum and SumMe datasets, respectively. Additionally, the inference stage is 300 times faster than our previously reported state-of-the-art method.

In this paper we propose a novel $Q$-learning algorithm allowing to solve distributionally robust Markov decision problems for which the ambiguity set of probability measures can be chosen arbitrarily as long as it comprises only a finite amount of measures. Therefore, our approach goes beyond the well-studied cases involving ambiguity sets of balls around some reference measure with the distance to reference measure being measured with respect to the Wasserstein distance or the Kullback--Leibler divergence. Hence, our approach allows the applicant to create ambiguity sets better tailored to her needs and to solve the associated robust Markov decision problem via a $Q$-learning algorithm whose convergence is guaranteed by our main result. Moreover, we showcase in several numerical experiments the tractability of our approach.

Deterministic policies are often preferred over stochastic ones when implemented on physical systems. They can prevent erratic and harmful behaviors while being easier to implement and interpret. However, in practice, exploration is largely performed by stochastic policies. First-order Bayesian Optimization (BO) methods offer a principled way of performing exploration using deterministic policies. This is done through a learned probabilistic model of the objective function and its gradient. Nonetheless, such approaches treat policy search as a black-box problem, and thus, neglect the reinforcement learning nature of the problem. In this work, we leverage the performance difference lemma to introduce a novel mean function for the probabilistic model. This results in augmenting BO methods with the action-value function. Hence, we call our method Augmented Bayesian Search~(ABS). Interestingly, this new mean function enhances the posterior gradient with the deterministic policy gradient, effectively bridging the gap between BO and policy gradient methods. The resulting algorithm combines the convenience of the direct policy search with the scalability of reinforcement learning. We validate ABS on high-dimensional locomotion problems and demonstrate competitive performance compared to existing direct policy search schemes.

Offline preference-based reinforcement learning (RL), which focuses on optimizing policies using human preferences between pairs of trajectory segments selected from an offline dataset, has emerged as a practical avenue for RL applications. Existing works rely on extracting step-wise reward signals from trajectory-wise preference annotations, assuming that preferences correlate with the cumulative Markovian rewards. However, such methods fail to capture the holistic perspective of data annotation: Humans often assess the desirability of a sequence of actions by considering the overall outcome rather than the immediate rewards. To address this challenge, we propose to model human preferences using rewards conditioned on future outcomes of the trajectory segments, i.e. the hindsight information. For downstream RL optimization, the reward of each step is calculated by marginalizing over possible future outcomes, the distribution of which is approximated by a variational auto-encoder trained using the offline dataset. Our proposed method, Hindsight Preference Learning (HPL), can facilitate credit assignment by taking full advantage of vast trajectory data available in massive unlabeled datasets. Comprehensive empirical studies demonstrate the benefits of HPL in delivering robust and advantageous rewards across various domains. Our code is publicly released at https://github.com/typoverflow/WiseRL.

Reinforcement learning has been applied in many fields, including robotics, video games, and recently Reinforcement Learning with Human Feedback (RLHF) [1, 2, 3, 4, 5], has become common in large language models. RLHF methods mitigate biases inherent in language models themselves [4, 6, 7]. Reinforcement learning models that address real-world problems predominantly utilize continuous models based on neural network architecture, known as Deep Reinforcement Learning (DRL). DRL methods typically involve a world model, agents interacting with the world, and a reward function that evaluates the effectiveness of actions based on the agent's policy towards predefined objectives [8]. Depending on whether the algorithm learns a specific world model, DRL algorithms are categorized into model-based DRL algorithms and model-free DRL algorithms [9]. Model-free DRL methods generally fall into three main categories: deep Q-learning methods [10, 11, 12], policy gradient methods [13, 14, 15], and actor-critic methods [16, 17, 18, 19]. Unlike model-based algorithms, model-free approaches circumvent model bias and offer greater generalizability, which contributes to their popularity in RLHF applications [4, 20]. Neural networks, which are the backbone of DRL methods, are associated with high computational costs and, therefore, resource intensive.

Deep Reinforcement Learning (DRL) has achieved remarkable success, ranging from complex computer games to real-world applications, showing the potential for intelligent agents capable of learning in dynamic environments. However, its application in real-world scenarios presents challenges, including the jerky problem, in which jerky trajectories not only compromise system safety but also increase power consumption and shorten the service life of robotic and autonomous systems. To address jerky actions, a method called conditioning for action policy smoothness (CAPS) was proposed by adding regularization terms to reduce the action changes. This paper further proposes a novel method, named Gradient-based CAPS (Grad-CAPS), that modifies CAPS by reducing the difference in the gradient of action and then uses displacement normalization to enable the agent to adapt to invariant action scales. Consequently, our method effectively reduces zigzagging action sequences while enhancing policy expressiveness and the adaptability of our method across diverse scenarios and environments. In the experiments, we integrated Grad-CAPS with different reinforcement learning algorithms and evaluated its performance on various robotic-related tasks in DeepMind Control Suite and OpenAI Gym environments. The results demonstrate that Grad-CAPS effectively improves performance while maintaining a comparable level of smoothness compared to CAPS and Vanilla agents.

Model-based Reinforcement Learning (RL) has shown its high training efficiency and capability of handling high-dimensional tasks. Regarding safety issues, safe model-based RL can achieve nearly zero-cost performance and effectively manage the trade-off between performance and safety. Nevertheless, prior works still pose safety challenges due to the online exploration in real-world deployment. To address this, some offline RL methods have emerged as solutions, which learn from a static dataset in a safe way by avoiding interactions with the environment. In this paper, we aim to further enhance safety during the deployment stage for vision-based robotic tasks by fine-tuning an offline-trained policy. We incorporate in-sample optimization, model-based policy expansion, and reachability guidance to construct a safe offline-to-online framework. Moreover, our method proves to improve the generalization of offline policy in unseen safety-constrained scenarios. Finally, the efficiency of our method is validated on simulation benchmarks with five vision-only tasks and a real robot by solving some deployment problems using limited data.

Algorithmic price collusion facilitated by artificial intelligence (AI) algorithms raises significant concerns. We examine how AI agents using Q-learning engage in tacit collusion in two-sided markets. Our experiments reveal that AI-driven platforms achieve higher collusion levels compared to Bertrand competition. Increased network externalities significantly enhance collusion, suggesting AI algorithms exploit them to maximize profits. Higher user heterogeneity or greater utility from outside options generally reduce collusion, while higher discount rates increase it. Tacit collusion remains feasible even at low discount rates. To mitigate collusive behavior and inform potential regulatory measures, we propose incorporating a penalty term in the Q-learning algorithm.